Frequently Asked Questions

Honest answer: often not, at least not the way you might think. Many processes benefit from simple automation long before AI enters the picture. Our job is to assess honestly where AI creates real value and where a simple workflow is the better choice. When in doubt: not every process needs AI.

Yes, absolutely. Most of our first conversations start with "we don't quite know yet what's possible." That's exactly why we begin with an assessment and show you which entry points exist in your company, before we talk about technology.

Yes, the initial conversation is free and without obligation. We typically spend 30 to 60 minutes understanding your situation and giving first impressions. Only when a concrete project emerges do we discuss effort and pricing.

Quite the opposite. Our focus is SMEs and public administration in the DACH region and Denmark. We believe mid-sized companies and government agencies have enormous potential for AI and automation, and deliberately need lean, pragmatic solutions, not enterprise projects.

We are deliberately tool-agnostic. Depending on the project we use n8n for workflow automation, Python for custom solutions, local LLMs (Llama, Mistral, Qwen via Ollama) for privacy-critical environments, and Claude or GPT where cloud APIs make sense. The tool mix follows the requirements, not the other way around.

OCR (Optical Character Recognition) handles printed standard text and is a commodity technology. ICR (Intelligent Character Recognition) uses AI methods and ensemble approaches, making it much more robust with difficult material. HTR (Handwritten Text Recognition) specializes in handwriting and requires its own models with real language understanding. In practice, all three are usually combined.

CER (Character Error Rate) measures the proportion of incorrectly recognized characters, WER (Word Error Rate) the proportion of incorrectly recognized words. These are the two standard metrics for text recognition. Modern AI pipelines achieve CER values below 2% on printed text and below 5% on handwriting, values that were considered unreachable only a few years ago.

An AI agent is a system that can independently research, make decisions, and execute tasks within your IT landscape, not just a chatbot that answers questions. The difference: an agent has a goal and a scope of action, a chatbot just has answers. Agents need clear boundaries, governance, and quality assurance.

RAG (Retrieval-Augmented Generation) connects a language model to a searchable knowledge base. Instead of relying solely on training data, the model pulls relevant documents from your data source at runtime and uses them for its response. This reduces hallucinations and ensures answers are based on your real, current data.

It depends on the data type. Personal data, health data, citizen data, and business-critical information should not flow uncontrolled into cloud services. That's why we prefer on-premise solutions where your data never leaves your infrastructure. For less sensitive data, cloud is often unproblematic, we decide that on a project basis.

On-premise means: the AI models and the entire processing pipeline run fully on your own infrastructure or in a data center you control. No API call to OpenAI, no data transfer to US providers, no Schrems-II risk. Your data stays where it belongs.

AI itself is neither compliant nor non-compliant, it depends on how it is deployed. Our solutions are GDPR-compliant by design: local processing, clear legal bases, transparent data flows, data processing agreements where needed. If required, we also help you with the documentation for your data protection officers.

Yes, and it's often more sensible than people think. Modern open-source models like Llama 3, Mistral, or Qwen run on reasonable hardware and deliver excellent results for many use cases. We advise you on hardware selection, setup (e.g. with Ollama or vLLM), and running them in a production environment.

Yes, but it always depends on your specific documents. Modern HTR pipelines achieve CER values below 5% for average handwriting, often significantly better. The key is that we train the pipeline on your real documents, not on lab test data. We find our honest answer in a proof of concept using your actual scans.

OZG, § 126 BGB (written form), the German BFSG and equivalents in Austria (AVG § 13) and Switzerland (VwVG Art. 21) oblige administrations to keep the paper channel open. This doesn't mean documents have to be manually typed, it means they have to be digitized and made processable. That's exactly what we build pipelines for.

Public administration (applications, building permits, objections), insurance (claims, legacy contracts), healthcare (handwritten notes, reports), archives and cultural institutions (historical documents), as well as industry and logistics (delivery notes, inspection records). Wherever paper is still a reality.

There's no one-size-fits-all answer, but we deliberately work in short cycles. A first analysis and strategy takes about 2-4 weeks. A proof of concept for a concrete use case usually 4-8 weeks. Production readiness then another 4-12 weeks, depending on complexity. We deliver early and often, so you quickly see whether the direction is right.

A proof of concept is a manageable prototype: the proverbial 80% solution. It answers two questions: does the approach work in principle? And does it achieve the quality you need? Only when both questions are clearly answered do we invest in full production readiness. This way you avoid expensive wrong decisions.

Vellmerk.ai is deliberately set up as a lean consultancy. For larger projects, we work with a network of experienced freelancers and specialists, we assemble the right team depending on the requirements. You get exactly the expertise your project needs, without overhead.

We work across industries in the DACH region (Germany, Austria, Switzerland) and Denmark. The consulting is flexible, remote, hybrid, or on-site, depending on the project and your preference.

It depends on the project. For consulting and workshops we usually work on a daily or hourly rate basis. For implementation projects we prefer fixed prices with clearly defined deliverables, so you have budget certainty. For long-term partnerships, retainer models are also possible.